Routing Bandwidth Guaranteed Paths with Local Restoration in Label Switched Networks Li Li, Milind M. Buddhikot, Chandra Chekuri, Katherine Guo
Abstract
setup policy and quality-of-service (QoS) constrained label switched paths (LSPs) between network nodes. The QoS constraint in the form of minimum or peak bandwidth guarantee per LSP has been considered most commonly in literature[3, 1, 10]. Such constraint-based routing [3, 5] is central to network traffic engineering [3, 2] and basic constructs of several new network services such as layer3 provider provisioned VPNs (PPVPN) [15] and layer-2 PPVPNs [4]. Specific examples of such constructs are the VPN tunnels in L3 PPVPNs, and the Virtual Private Wire Service (VPWS) and Virtual Private LAN Service (VPLS) in L2 PPVPNs. The two main steps in setting up LSPs are (1) computing a path that satisfies required constraints and (2) establishing and maintaining forwarding state along that path. Clearly, the routing algorithm used in step 1 is a basic building block for new network services.
The emerging Multi-Protocol Label Switching (MPLS) networks enable network service providers to route bandwidth guaranteed paths between customer sites [3, 2, 8, 5]. This basic Label Switched Path (LSP) routing is often enhanced using restoration routing which sets up alternate LSPs to guarantee uninterrupted connectivity in case network links or nodes along primary path fail. In this paper, we address the problem of distributed routing of restoration paths, which can be defined as follows: given a request for a bandwidth guaranteed LSP between two nodes, find a primary LSP and a set of backup LSPs that protect the links along the primary LSP. A routing algorithm that computes these paths must optimize the restoration latency and the amount of bandwidth used. In this paper, we introduce the concept of “backtracking” to bound the restoration latency. We consider three different cases characterized by a parameter called back), (2) limtracking distance D: (1) no backtracking (D ited backtracking (D k ), and (3) unlimited backtracking (D 1). We use a link cost model that captures bandwidth sharing among links using various types of aggregate link state information. We first show that joint optimization of primary and backup paths is NP-hard in all cases. We then consider algorithms that compute primary and backup paths in two separate steps. Using link cost metrics that capture bandwidth sharing, we devise heuristics for each case. Our simulation study shows that these algorithms offer a way to tradeoff bandwidth to meet a range of restoration latency requirements.
=
=
Note that the failure of nodes or links along LSPs leads to service disruption. The NSPs have considered enhancing the reliability and availability of new services using restoration routing, which sets up alternate paths to carry traffic under fault conditions. There are two variants of restoration routing: (1) End-to-end restoration: routes two link disjoint paths – one primary path and one backup path between every source, destination node pair [9, 13]. Resources are always reserved on the backup and are guaranteed to be available when the backup path is activated. In the event of a failure along the primary path, the source node detects path failure and activates the backup path. In the absence of an explicit signaling protocol, source node learns of link failures via intra-domain routing updates such as OSPF link state advertisements (LSA) packets or IS-IS Link State Packets (IS-IS LSP) which are processed in the slow-path typically in a routing daemon in the router or switch controller. Also, such packets are typically generated by timer driven events. This causes very large restoration latencies of the order of seconds or at best 100s of milliseconds. (2) Local restoration: routes a primary path between the source and destination and a set of paths that protect the links along the primary path [11]. It can minimize the need for source node to be involved in the path restoration and therefore can achieve fast restoration. Though local restoration may use
=0
1 Introduction The emerging Multi-Protocol Label Switching (MPLS) networks enable network service providers (NSPs) to Li Li, Milind M. Buddhikot, Chandra Chekuri, Katherine Guo are with Bell Labs, Lucent Technologies, NJ USA (e-mail: ferranlli,mbuddhikot,
[email protected],
[email protected]).
1 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
network and n is the number of nodes in the network. In contrast, our running time is O hnlogn hm where h is the primary path length for D < 1 and a constant for D 1. Their algorithm does not explicitly encourage link sharing among multiple backup paths for the same primary path and therefore, in the worst case may compute backup paths that are largely link disjoint. Their work also does not provide a theoretical and algorithmic framework for local restoration that allows a network designer to tradeoff bandwidth for better restoration latency. Our research reported here overcomes these limitations.
more (bandwidth) resources, it is attractive as it can meet stringent restoration latency requirements that are often of the order of 50 ms – similar to existing SONET protection mechanisms [18]. The problem of end-to-end and local restoration routing has been addressed in several recent research papers [9, 13]. In this paper, we primarily focus on the problem of local restoration routing. We first describe the new concept of restoration routing with backtracking characterized by a single parameter called backtracking distance D. We consider three cases: (1) no backtracking (D ), (2) limited backtracking (D k ), and (3) unlimited backtrack1). We first consider the joint optimization ing (D of primary and backup paths and show that the problem is NP-hard. We therefore propose computation of primary and and backup paths in two steps. In the case of D D k , we show that even the two-step optimization prob1 an optimal solution lem is NP-hard, whereas for D exists for computation of backup path. We then describe heuristics that use per-link network state in the form of (1) residual link capacity (Rl ), (2) bandwidth consumed by primary paths (Fl ), (3) bandwidth consumed by backup paths (Al ) and optionally a fixed sized Backup Load Distribution matrix [9, 13]. We also present simulation results to characterize the performance of these algorithms. We show that our algorithms offer a means to tradeoff bandwidth to meet a range of restoration latency requirements.
)
1.2 Outline of the Paper The outline of the rest of this paper is as follows: Section 2 presents the background material for the remaining discussion in the paper. Section 3 introduces the concept of restoration routing with backtracking and describes the three cases we consider. In Section 4, we show that the joint optimization of primary path and corresponding local restoration subgraph is NP-hard and advocate a two-step approach for computing them. Section 5 describes in detail backup subgraph computation algorithms for D ; k; 1 cases. The concept of post-processing to achieve further bandwidth savings and the post-processing algorithms for each case are discussed in Section 6. Section 7 describes our simulation experiments in detail. Finally, Section 8 presents our conclusions and future directions.
=0
=
+
=
=0
=
=
(
=
=0
1.1 Related Work
2 Background The MPLS Forum has proposed use of constraint-based routing to setup “protection LSPs” to bypass failed links and nodes [3]. In this scheme, the upstream router detects the link failure and tunnels all traffic on that link along a preestablished MPLS LSP implemented using the label stacking mechanism. Clearly, this mechanism can be extended to selectively reroute certain LSPs on the link instead of the entire link traffic. Extensions have been proposed to the RSVP resource reservation protocol to signal such a protection LSP [7]. However, the existing fast reroute model does not require protection LSPs to have bandwidth guarantees and aims to provide continuation of best-effort connectivity. The orthogonal question of how to route such best-effort or bandwidth guaranteed protection LSPs is still a topic of investigation. Our work provides an algorithmic framework and candidate algorithms to solve this important problem. The paper by Kodialam et al.[11] represents the current state-of-the art approach for local restoration. Their algorithm iteratively uses a shortest path algorithm to find restoration path for each link in the primary path. The shortest path algorithm is invoked for each edge in the network. Therefore, the running time of their algorithm is O mnlogn m2 where m is the number of edges in the
(
In this section, we will present relevant background material on various aspects of the problem of routing bandwidth guaranteed backup paths.
2.1 Network and Service Model The label switched network is modeled as a graph G = (V; E ), jV j = n, jE j = m where graph nodes correspond to
the network nodes and graph edges correspond to the physical links between the nodes. Each link in fact consists of two simplex links in opposite direction: one for the transmit path and the other for the receive path. The graph nodes can be classified into two types[3]: (1) Label Edge Router (LER): These are nodes at which the network traffic enters or leaves the network. Such nodes are connected to the customer edge (CE) routers which use the transport services of the LSP service provider, (2) Label Switched Router (LSR): these are the transit routers that forward packets based on the MPLS labels. Each LER independently receives requests from connected CEs to setup bandwidth guaranteed paths. Each such s; d; b where s is the request r is modeled as a 3-tuple r
+ )
=(
2 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
)
( )
such partial information scenario, referred to as the O jE j information case here on, wherein for every link l in netV; E with capacity Cl , three state variables are work G maintained and exchanged among peering routers: (1) Fl : Amount of bandwidth used on link l by all primary paths that use link l. (2) Al : Amount of bandwidth reserved by all backup paths that contain link l. (3) Rl : Residual caFl Al . Norden pacity on the link l defined as Cl et al.[13] proposed algorithms that use these three per-link state variables and a new form of state called Backup Load Distribution (BLD) matrix. Specifically, given a network with m links, each network node (router or switch) maintains a m m BLD matrix. If the primary load Fj on a link j is b units, entries BLDM i; j ; i m; j 6 i, record what fraction of b is backed up on link i. Note that this approach, which we call as the O jE j2 information case, allows maximum bandwidth sharing. The algorithms we propose in this paper work with both cases of state information.
source node, d is the destination node and b is the bandwidth requirement [4]. In MPLS networks, an LSP between s and d is a simplex flow, that is, packets flow in one direction from s to d along a constrained routed path [3]. For reverse traffic flow, additional simplex LSP must be computed and routed from d to s. Clearly, the path from s to d can be different from the path from d to s. Also, the amount of bandwidth reserved on each paths can be different. This request model is often referred to as pipe model in the VPN literature [3]. We call this model and corresponding constrained path routing asymmetric request model. The algorithms reported in this paper assume this request model.
=(
)
( + )
[ ]1 (
2.2 Fault Model In the context of protected path routing it is important to consider two kinds of failures, namely link failures and router failures. A common fault model for link failures assumed in the literature and justified by network measurements is that at any given time only one link in the network fails [18]. In other words, in the event of a link failure, no other link fails until the failed link is repaired. We consider single link failure in this paper. We do not address router failure case due to space constraints.
=
)
2.5 Modeling Link Cost Each
link
(interchangeably
called
edge)
has
C ; R ; F ; A state variables associated with it and l
l
l
l
may be assigned one or more link costs. In response to the s; d; b , the source node uses its knowltunnel request r edge of network state such as topology, Cl ; Rl ; Al ; Fl for all links l 2 E and optionally BLD matrix to compute two things: (1) a primary path P s u1; u2 ; d uk where edge li ui ; ui+1 2 E , (2) a subgraph Gsub Vsub ; Esub such that a backup path exists in Gsub that can route traffic from s to d if any link li fails. The computation of P and Gsub is based on two kinds of link costs described below. Cost of link l in the primary path: If a link is used in s; d; b , then b units of the primary path P of request r bandwidth has to be consumed on each link in P . Therefore the cost l of using a link l in the primary path is b if Rl b, otherwise 1. Cost of link l in backup path: Consider a primary path P with m links ; ; :::m . Let us consider a representative link l that is a candidate link in a backup path LBi for link i 2 P . We are interested in characterizing the cost of using l. The amount of bandwidth that can be used on l for free to backup i is F R l; i Al Backup load induced by i on l before request r. In the O jE j information case, free bandwidth is F R l; i Al Fi , whereas in the O jE j2 information case, the free bandwidth on l is F R l; i Al BLDM l; i . For the rest of the paper we focus only on the O jE j2 information case. However, all our results apply equally to the O jE j information case. From the perspective of backup routing, every link has
=(
2.3 Bandwidth Sharing We distinguish bandwidth sharing into two categories: (1) Inter-request sharing: The single fault model dictates that two link-disjoint primary LSPs corresponding to two requests each of b units do not fail simultaneously. This allows them to share a single backup path of b units. In other words, inter-request bandwidth sharing allows one of the two primary paths to use the backup links “for free.” (2) Intra-request sharing: In local restoration, since each link on the primary path requires a backup path, and only one link failure can happen at any given time, the backup paths for different links in the same primary path can share links. Also, backup paths can share primary path links. Notice this type of sharing does not exist in the end-to-end path restoration schemes. Our goal is to develop online distributed local restoration routing algorithms that utilize both inter-request sharing and intra-request sharing in order to minimize the total bandwidth reserved.
=(
)
=( = )
= ( )
= )
=(
(1 2
( )
2.4 Network State Information In our work we focus on online routing algorithms that route a new path request based only on the knowledge of current state of the network and do not exploit properties of future requests. One can design schemes with varying degrees of partial state. Kodialam et al. [10] describes one 3
Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
)
)
( ) [ ]=
[ ] = [ ] =
(
[ ] )
( )
Table 1. Vector and scalar link costs Vector W of costs for link l
Simplified scalar cost for link l
8> 1 < bC w[l; i] = >: F R[l; i] C + (b F R[l; i]) C F
F
R
[ ] [ ] [ ] [ ]
if b F R l; i > Rl if b F R l; i (1) if b > F R l; i and b F R l; i < Rl
(
l
= =
i
l
l
(
F l
Naïve local restoration Local Backup LB2 LB1
R
s
u1
)
Local restoration with backtracking
u3
s
u0
u1
LB1
v2
v1
v4
v3 u2
b LB4
LB3
v2
v1
u2
u3
d
d Downstream
Primary
LB2
Figure 1. Local restoration and backtracking to address this by “localizing” fault detection and subsequent restoration actions. Consider the simple example shown in Figure 1-(a) where there are as many backup paths LBi as links in the primary path and restoration is truly local and rapid. Under single link failure model, completely link disjoint LBi s result in bandwidth wastage. Instead, a model wherein backup paths (1) share links amongst them and also (2) share links in the primary path can be used to minimize bandwidth usage at the cost of increased restoration latency. Consider the scenario shown in Figure 1-(b): In this case, links u1 ; u2 , and u2 ; u3 are protected by a single backup path, LB1 u1 ; v1 ; v2 ; u3 . Clearly, in a na´ive setup, the backup path for link u2 ; u3 is u2 ; u1 ; LB1 , which backtracks over primary path by a distance of one link namely u2 ; u1 . Similarly for link u3 ; d , the backup path u3 ; u2 , u2 ; u1 , u1 ; u0 , LB2 , backtracks by a distance of three links. If restoration must be truly local, i.e failure of u2 ; u3 must be restored by u2 and failure of u3 ; d must be restored by u3 , then u2 , u3 must “switch back or backtrack” traffic on links u3 ; u2 , u2 ; u1 , u1 ; u0 . Such repeated link traversals cause undesirable traffic loops, loss and bandwidth wastage. This is especially true when the node performing the restoration is multiple hops away from the node to which failed link belongs. Alternatively, if nodes u2 and u3 can inform u1 outof-band of their links failures, u1 can perform the restoration. However, in this case restoration is non-local and requires new form of signaling or modifications to existing routing protocol state updates. We define the backtracking distance D as the maximum number of hops (links) in the primary path that must be
[ ]
(
(
E P
i
l
a
min 2 ( ) F R[l; i] (3) A max 2 ( ) BLDM [l; i] where E (P ) denotes the edge set used in path P and V (P ) denote the node set used in path P . Correspondingly, the cost w of using link l on the backup path for any link in P FR
if b F Rl > Rl if b F Rl (2) if b > F Rl and b F Rl < Rl
F
lead to long restoration latencies. Local restoration attempts
[ ]
[ 1] [ 2]
w
)
two kinds of bandwidth available: (1) Free bandwidth (F R): that is completely sharable and does not require extra resource reservation. (2) Residual bandwidth (R): is the actual capacity left unused on the link. If the LSP request size b > F Rl , then b F Rl units of bandwidth must be allocated on the link to account for the worst case backup load on the link. If the residual bandwidth Rl falls short of b F Rl (i.e b F Rl > Rl ), then the link l cannot be used on the backup path and is called an “infeasible link”. Given this, the cost of using link l on a backup path to backup link i on primary path consists of two parts: (1) cost of using the free bandwidth on the link and (2) cost of using the residual bandwidth on the link. Equation 1 illustrates the exact form of cost w l; i incurred to backup link i on l. The cost metrics CF (CR ) should be selected in such a way that selecting a link with high residual capacity Rl , results in smaller cost. See [13] for more details on these cost metrics. Since there are m links in the primary path, each candidate link l has m associated costs w l; ; w l; w l; m . Given that a link l may be used in backup paths for multiple links i in primary path, to make the problem tractable, we have to follow a pessimistic approach and reduce the free available bandwidth for l as follows:
8> 1 < = > bFRC C + : (b F R ) C
E P
(
l
)
(
)
(
can be computed using Equation 2.
(
3 Concept of Local Restoration with Backtracking In this section, we introduce the concept of backtracking and discuss properties of backup subgraph Gsub in various cases of backtracking.
3.1 Backtracking In the case of end-to-end restoration routing, detection of link failures and subsequent restoration by source node can
)(
)
)
4 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
)
(
)( ( )
( =(
)
)
(
)(
) )
)
(
)
traversed upstream towards the source before reaching the node at which the required backup (restoration) path originates. Three cases that are of interest are as follows:
backup path. (2) Equation 1 shows that the cost of a backup link also depends on which set of primary links it protects. The following theorem characterizes the complexity of this optimization problem.
=0
No backtracking (D ): In this case, the backup path must originate at the node at which the failed link originates. This case represents true local restoration and provides best restoration latency. The link restoration paths computed in this case can use a subset of primary path links downstream towards the destination node.
0
Theorem 4.1 For all values of backtracking D ( ; k; 1), computation of primary path P and a restoration subgraph Gsub such that they are jointly optimal is NP-hard. The proof of this theorem can be found in [12].
5 Two-step Algorithms
=
Bounded backtracking (D k ): In this case, the backup path can originate at a node on the primary path up to k hops away from the node that owns the link.
In this section, we consider two-step algorithms that compute the primary path in the first step and then construct the restoration graph Gsub in the second step. The basic pseudo-code for these algorithms is presented in Table 2.
=
Infinite backtracking (D 1): This case allows unlimited backtracking and therefore, in the worst case may result in backup paths that originate at the source node. The end-to-end restoration can be considered a special case of this where the maximum backtracking allowed is equal to length of the primary path but the restoration always has to be initiated at the source node.
=0
Table 2. A two-step algorithm 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15.
may require the highest amount backup Clearly, D bandwidth, and lowest restoration latency whereas D 1 requires the least amount of bandwidth and highest restorak is interesting as it altion latency. Therefore, case D lows us to tradeoff bandwidth for better restoration latency. In the rest of the paper we address the problem of computing primary path P and its corresponding backup restoration graph Gsub for various cases of backtracking.
=
=
In this section, we consider the problem of joint optimization of computing primary path and its associated backup subgraph Gsub such that the total cost is minis; d; b and a given network mized. For a new request r G V; E , ideally we would like to have a primary path P s u1 ; u2; d uk and Gsub Vsub ; Esub such that the total cost l2E (P ) l wl is minl2Esub imized under the following constraints: for every edge l ui ; ui+1 2 E P , there is a path from uj to ut in Gsub such that i j D and t i , i.e. the backtracking distance of the backup path does not exceed D hops and the path must end at a node that is downstream from ui+1 in P . There are two characteristics of this joint optimization problem that make it hard to solve: (1) For each edge, its cost depends on whether it is used for primary path or
=( P= )
=(
)
0
( )
)
+P
=(
G G
1
G G
G
The problem of primary path routing – computing a bandwidth guaranteed path between a pair of source and destination nodes- has received significant attention in recent years. The simplest solution to this problem is to use Dijkstra’s shortest path algorithm. However, its drawback is that though it yields an optimal solution for a single request, over a span of multiple requests, it can lead to high request rejection and low network utilization [1, 10]. Several new algorithms such as Widest Shortest Path First (WSPF) [1], Minimum Interference Routing [10], rectify this limitation. The WSPF algorithm alleviates this problem by selecting the shortest path with maximum (“widest”) residual capacity on its component links. We choose WSPF (line 2) as a good tradeoff between simplicity and performance. In the remaining section, we discuss backup path com(line 6), D 1 (line 9), and putation for cases of D D k (line 12) in detail.
4 Joint Optimization of Primary and Backup Paths
=( ) =( =
Given graph G and req r=(s,d,r); P=WSPF(G,r); // Primary path using WSPF If (P==NULL) return reject; Switch (D) // Backup path computation case 0: sub =modified_directed_Steiner(G,P,r); break; : case sub =Suurballe(G,P,r); break; case k: sub =modified_Suurballe(G,P,r,k); if ( sub ==NULL) return reject; postprocessing( sub); return accept;
)
+1
=
=0
=
5.1 Computing Backup Path for D = 0 In this case, we need to find a subgraph Gsub such that there is an alternative path with no backtracking if any link 5
Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
in the primary path fails, and the total link costs for all backup paths are minimized. The following theorem states that the problem is NP-hard. Primary Path P
A s u1
u2
u3
B u4
u6 d
u5
primary path, there are multiple restoration paths that end on the immediate neighbor or neighbors downstream. This encourages sharing of primary path links and minimizes new bandwidth reservation. In our example (Figure 2), in the case of links u3 ; u4 and u4 ; u5 , the backup paths u3 ; v3 ; v4 ; u5 , and u4 ; v4 ; u5 share the primary path link u5 ; u6 for restoration. (2) Also, the explicit tree structure ensures that local restoration paths share links. In our exu1 ; v2 ; u3 ample, restoration path for u1 ; u2 is LB which is also part of the restoration path u2 ; v1 ; u1 ; v2 ; u3 for u2 ; u3 . In the Section 6, we describe an algorithm that optimally selects restoration paths from the Gsub .
Directed Steiner tree Gsteiner rooted at u6 (d) with terminal nodes u1 u2 u3 u4 u5 d u6
( (
v5 Backup sub-graph Gsub with Steiner tree edges reversed
C
v4
v3
v1
u2
u3
u4
u5
(
v3
u4 s u1
u3
d
u6
v2 v1
v2
=(
=0
)
For proof, please refer to [12]. This theorem also gives us the algorithm for the computation of Gsub . Specifically, (1) Compute a directed Steiner tree GStei on subgraph G0 V; E E P , rooted at d and with the nodes V P fdg in the primary path as its terminal nodes. (2) Reverse all the edges of the tree to obtain backup subgraph Gsub . Figure 2 illustrates this with an example. It shows the primary path P (s u1 , u2 , u3 , u4 , u5 , d u6 ), the corresponding Steiner tree Gstei and the backup subgraph Gsub which has links that are reverse of those in Gstei . Among the approximation algorithms that compute directed Steiner tree such as [17, 14, 6], we choose the SCTF algorithms [14]. SCTF has good approximation ratio if the graph is not very asymmetric (asymmetry is measured as the sum of the larger cost of edges u; v and v; u divided by the sum of the smaller cost.). Applying SCTF algorithm with minor modifications, the backup subgraph is computed as follows: the current subgraph Gsub is initialized to be the single node d; the algorithm then iteratively computes a shortest path from the current set of nodes in the primary path that are not in the current subgraph to the current subgraph; the algorithm terminates when all the primary nodes are in Gsub . Each iteration involves computing a shortest path from one set of nodes to another set of nodes. This requires only one invocation of Dijkstra’s algorithm by adding dummy source and destination. The running time is O hnlogn hm where h jV P j ,n jV j and m jE j. The reversed directed Steiner tree as a restoration subgraph Gsub has several advantages: (1) For every link in the
( )
= =
( =
+
)
= ( ) 1
)
)
=(
(
(
)
=
=
) )
)
) ) ( )( ) Computing Backup Paths for D = k (
=
( )
)
=(
=0
( ))
( )
Given a primary path P for request s; d; b in directed V; E , if there is no constraint on the backtrackgraph G ing distance, the following procedure due to Suurballe [16] computes a minimum cost backup set of edges for P . The idea is to reverse the direction of the edges on the path P and set their cost to zero. All other edges are assigned cost as defined in Equation 2. Compute a shortest path Q from s to d in the resulting graph G0 . Suurballe [16] shows that the edges of Q that are not in P represent the optimal backup path for the primary path P with D 1. Figure 3-(a) illustrates an example, where primary path P is u1 ; : : : ; u10 and backup subgraph consists of paths u1 ; v1 ; v2 v3 ; u3 , u2 ; v4 ; v5 ; v6 , v7 ; u4 ; v8 ; v9 ; v10 ; u6 , u5 ; v11 ; v12 ; v13 ; u10 .
Theorem 5.1 Given a network G V; E and a primary path P from s to d, the problem of computing optimal case is NP-hard. backup subgraph Gsub for D
(
)
5.2 Computing Backup Paths for D = 1
u1 (s)
u2
Figure 2. Example of backup path for D
)
(
(
u5 v4
v5
(
)
5.3
(
For this case, we state the following theorem.
=( =
)
Theorem 5.2 Given a primary path P in G V; E , computing the minimum-cost backup subgraph Gsub based on cost wl using bounded backtracking of D k is NP-hard.
( )
For proof, please refer to [12]. For the computation of k case, we build upon the probackup subgraph for D cedure due to [16]. Table 3 illustrates the pseudo code of our heuristic algorithm. We first compute a backup path Q for P as in Section 5.2. It is easy to see that the cost of Q is a lower bound on the cost of any solution for the D k case. The edges of Q allow the edges on the primary path to be restored, however the maximum distance some edge in the primary path has to backtrack might not be bounded by k . Procedure maxD checks the maximum backtracking distance and if it exceeds k , additional paths are added in a greedy fashion to the edges of Q. We use an example to illustrate our ideas. Consider the example in Figure 3(a) after line 2 in Table 3. We use k in the example. This requires that the restoration path for a link ui ; ui+1 in the primary path must originate at ui or ui 1 . Note that
=
=
=
=1
6 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
(
)
v1 S=u1
v2
v3
u2
v9
v8 u4
u3
u5
v4
u6
u7
u8
u9
u10= d
(a) Example of primary and backup path for
u2
D=1
v5
v13
the edges of Q computed from Suurballe’s procedure allow all edges except u7 ; u8 , u8 ; u9 and u9 ; u10 to have a restoration path with backtracking distance at most . In the example, the backup path with the smallest backtracking distance for the link u8 ; u9 is the one that originates at u5 which has a backtracking distance . If we add paths BP1 ; BP2 as shown in Figure 3-(b) to the Gsub , we can obtain an efficient solution that satisfies the backtracking distance constraint. Our algorithm (Line ) finds such paths in the following way. For every link ui ; ui+1 in the primary path which does not have a restoration path with backtracking distance bounded by k , we add a path from some node in the set fui k ; ui k+1 ; : : : ; ui g to reach a node in fui+1; ui+2 ; : : : ; dg. This ensures that we satisfy the requirement for all links on the primary path. We process unsatisfied links in the order of their increasing distance to the destination. In the example, the link u9 ; u10 is considered first. To satisfy this link we consider adding paths from either u8 or u9 to reach the node u10 . We consider u8 first and if we are unable to find a path to reach u10 , we then consider u9 . In general, when trying to satisfy link ui ; ui+1 , we search for paths starting with ui k and stop when we find a path. In the example, we find BP1 from u8 . Note that we find a shortest path at each step to minimize the cost of solution. Once a link is satisfied, we move to the next unsatisfied link (farther away from the destination) and
(
)
(
)
1
3
u10= d
v13
v12
= 0 and D = k (
)
)
(
)
In the following we first discuss the concept of postprocessing and provide algorithms for each case of backtracking.
6.1 Need for Post-processing
)
(
(
v11
u9
BP1
6 Post-processing
5
(
u8
u7 BP2
=
[G
)
u6
D=k
(
modified_Suurballe(G,P,r,k) 1. Q=Suurballe(G,P); 2. If (Q==NULL) return NULL; 3. 4. if (maxD(P,Q)>k) 5. a =compExtraPaths(G,P,Q,k); 6. if ( a ==NULL) return NULL; 7. else sub =Q a ; 8. return sub ;
)(
u5
repeat the above procedure. In the above example, adding path BP1 satisfies both u9 ; u10 and u8 ; u9 and the next unsatisfied link is u7 ; u8 . The process stops when no unsatisfied links remain. All the BPi so obtained combined with the original Q – i.e. Gsub fBPi g [ Q provides the restoration graph that satisfies required backtracking constraint. In the above procedure, we need to find a shortest path from a node u 2 fui k ; : : : ; ui g to reach any node in fui+1 ; ui+2 ; : : : ; dg. This can be implemented by Dijkstra’s algorithm by first shrinking the nodes fui+1 ; ui+2 ; : : : ; dg into a single super node. In computing the shortest path we reduce the cost of the edges in Gsub computed till now to zero so as to make sure that we reuse the edges that are already in our solution as much as possible.
=
(
v10
v6
(b) Basic ideas in
Table 3. Heuristic Algorithm for Backup Path Computation of D k case
G G G
u4
u3
Figure 3. Computing restoration paths for D
G
v9
v8
v7
v12
v6
s=u1
v3
v4
v11
v7 v5
v2
v1
v10
v2 v1
v4
v3
6
LB2
LB1
)
s=uo
1
u1
2
u2
3
u3
4
5
u4
d=u5
Figure 4. Need for post-processing Figure 4 illustrates primary path P (u0 ; u1 ; u2 , u3 ; u4 , u5 ), and the corresponding backup subgraph G for D = 0 for a request of size b = 18 units. Recall that G is the reverse of the Steiner tree rooted at u5 . Here link 1 (u0 ; u1 ) is backed up using path LB1 , constituted by (u0 ; v1 ), (v1 ; v2 ), (v2 ; v4 ) (v4 ; u5 ), whereas link
)
sub
sub
7 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
4 (u ; u ) is backed up by using path LB constituted by (u ; v ), (v ; v ), (v ; u ). Consider link 6 (u ; v ) on LB , with residual bandwidth R = 22, backup load A = 32. The backup load induced by link 4 on link 6 is BLDM [6; 4] = 20, whereas load induced by link 1 on link 6 is BLDM [6; 1] = 10. 3
4
3
3
(u ; u ) on the path P , the restoration path has to origi-
2
3
4
4
5
0
1
i
6
(
6
4
6 6 (
4
6
)
+1
Unfortunately, when link cost wl is assigned for Steiner tree computation, a pessimistic assumption has to be made: units induced the worst case backup load on link is by link . This has to be done to account for possibility that backup path for link may include link . So the extra bandwidth reservation on link is b A6 BLDM ; or (18 - (32 -20))=6 units. In other words, the Equation 2 we employ to compute F Rl in Steiner tree computation is rather conservative as it assumes that any candidate link in backup graph may have to backup worst case load from a link in the primary path. If we introduce post-processing that follows computation of Steiner tree and path selection, we can fix this situation. Specifically, in our example, if we account for the fact that only backup path LB1 for link uses link , we can see that the amount of free bandwidth on link is A6 BLDM ; , which is greater than the request size. Clearly, in that case no extra bandwidth has to be reserved on link . Such post processing of the solution can reduce extra bandwidth reservation for backup paths and therefore save bandwidth. We provide post-processing algorithms for each case of D. The two main goals of these algorithms are: (1) Identify for each primary path link, a set of potential restoration paths with minimum backtracking distance. (2) Minimize the total bandwidth reserved on the backup paths subject to assignment of restoration paths. We state the following theorem for post-processing for all three cases of backtracking:
6 20 6 (
i+1
nate at ui . In Gstei , for every ui in P , there is a unique directed path LBi from ui to the destination d. We could use this path LBi as a restoration path for ui ; ui+1 , however we notice that LBi can intersect the path P at several places that and in particular there could be a vertex uj , j i lies on LBi . In this case, it is easy to see that the portion of LBi from ui to uj is sufficient to restore the link ui ; ui+1 . In Figure 2, the path LB3 is u3 ; v3 ; v4 ; u5 ; v5 ; u6 . However the portion of LB3 that is u3 ; v3 ; v4 ; u5 is sufficient to restore the link u3 ; u4 . More generally, for the unique path LBi from ui to d, let uj be the first vertex in LBi such that j i . Let Q0i be the portion of LBi from ui to 0 uj . We use Qi to restore the link ui ; ui+1 . In Figure 2, the path LB2 intersects P at u1 ; u3 ; u5 and finally reaches d u6 . We use the portion of LB2 from u1 to u3 , hence Q02 is u2 ; v1 ; u1 ; v2 ; u3 . It is easily seen that this Q0i as constructed above is necessary and sufficient for the restoration of ui ; ui+1 . Now we address the issue of minimizing the bandwidth. For each link ` in Gstei , let S` be the set of all i such that the link ` is in the restoration path Q0i for link ui ; ui+1 . In our example from Figure 2, the link v2 ; u3 is used by Q01 and Q02 , hence S(v2 ;u3 ) f ; g. For link v4 ; u5 , we have S(v4 ;u5 ) f ; g. In other words, S` is the precise set of all primary path links that use ` in their backup paths. Therefore it is sufficient to reserve bandwidth on ` that will satisfy the links in S` and not all the links of P . This results in savings in bandwidth. More precisely, the free bandwidth on ` for backing up links in S` is F R` i2S` F R `; i . This is to be contrasted with the pessimistic estimation F R` i2E (P ) F R `; i used in the computation of Gstei . Since S` is a proper subset of E P , we can potentially use more free bandwidth on ` than estimated. The revised bounds on the free bandwidth can be easily computed once the restoration paths Q0i are computed for all i. However, we note that both the Q0i for all links of P and the F R` for all the links in Gstei can be computed in a single DFS traversal of Gstei by keeping some simple state information at the vertices of the trees. The traversal of Gstei takes O m time where m is the number of vertices (edges) in Gstei . The explicit enumeration of the restoration paths needs an additional O h time, where h is the sum total of the lengths of Q0i . Hence, the running time is linear in the size of the tree and the size of the output.
1
(
(
[6 4])
)
(
(
+1
= ( (
1
[6 1]) = 32 10 = 22 6
)
(
)
)
)
= 34
= 12
(
(
)
(
= min
= min
Theorem 6.1 Given a primary path P , the backup subgraph Gsub that our algorithms compute is a forest consisting a set of trees fTi g. The amount of bandwidth needed to be reserved can be computed optimally by a single traversal of P and one ordered traversal of each tree Ti . The backup paths for each primary edge can be found in time linear in the output size.
)
)
[ ]
)
)
[ ]
( )
( )
()
We do not give a formal proof of the theorem. However the description of the algorithms in the following subsections contains all the ideas behind the proof.
6.2 Post-processing for D = 0
6.3 Post-processing for D = 1; k
For this case we describe a post-processing algorithm that results in the minimum use of bandwidth subject to using the edges of Gsub for restoration. We use the example from Figure 2 to illustrate the algorithm. The algorithm is based on the fact that the backup edges form a Steiner tree Gstei on the vertices of the primary path P . For a link
= =0
=
The post-processing algorithms for D 1 and D k are conceptually similar to the case D and hence we omit the details. However we do need to take care of two issues. First, for a given link ui ; ui+1 in P there might be several restoration paths that satisfy the backtracking con-
(
8 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
)
straint. By choosing the restoration path with the minimum backtracking distance we can make the restoration path for each link unique. Second, instead of a single tree as in the D case we might have a forest consisting of many trees. We process each of these trees much as we do in the D case.
=0
process. We use 6 quantization levels for the EWSPF algorithm which are : ; : ; : ; : ; : and : times the maximum requested bandwidth [13]. The network topology we use is the same as the homogenous topology in [13]. The topology represents the Delaunay triangulation for the 20 largest metro areas in continental U.S. All links are of the same capacity (OC-48). Additional results for different topologies can be found in [12].
0 05 0 1 0 3 0 5 0 7
=0
7 Simulation Experiments
7.2 Performance Metrics
In this section, we describe simulation experiments that characterize the benefits of our proposed local link restoration schemes over existing end-to-end path restoration schemes. We conduct a set of experiments that com; k; 1) with two end-to-end pare our three schemes (D restoration schemes, namely Enhanced Widest-ShortestPath First(EWSPF) [13], and simple Shortest Path First (SPF). EWSPF exploits bandwidth sharing among backup paths and has been shown to perform better than other schemes in the literature [13]. The SPF scheme uses link costs based on the residual capacity and computes two independent paths: one used as primary and the other as backup and do not attempt to share backup paths. It is used as a base line to show how much benefit bandwidth sharing schemes can provide. Because our topology is relatively small, we case for the limited backtracking alonly consider D gorithm.
We defined performance metrics to characterize (1) restoration latency and (2) bandwidth sharing performance [13].
=0
7.2.1 Restoration Latency Performance A brute force direct simulation of faults and subsequent measurement of restoration latency requires a very complex model that simulates multi-gigabit data traffic, link faults and corresponding fault propagation traffic in a complete network. Such a simulation that involves simulation of traffic instead of just request arrivals and corresponding route computations yields at best only representative performance numbers. So instead of taking this inordinately complex approach, we use an indirect way to measure restoration latency based on the following two performance metrics: (1) Histogram of Backtracking Distance (HBD): Given a primary path of length l, the restoration model used dictates the amount of worst case backtracking for a link in the primary path. For example, with local restoration and back, the backup path may backtrack tracking distance of D ; or links. On the other hand for end-to-end restoration, for the st link in the path, backtracking is equal to zero, whereas for the lth link in the path, restoration backtracks to source node and has backtracking distance of l . The more the backtracking, the more is the restoration latency. Therefore, we compute a histogram, where bin i corresponds to total number of links in the admitted primary paths for which backup path backtracks i links. Clearly, this histogram characterizes the probability that a given amount of backtracking will occur. (2) Average Case Backtracking Distance (ABD): We define this metric as follows:
=1
7.1 Simulation Setup This section describes the network topology, traffic parameters and performance metrics used in the simulation.
01 2
Table 4. Simulation parameters Parameter Request (REQ) arrival Call holding time (HT) Simulation time (STT) Request Volume (RV) during entire STT Max single LSP REQ size Mean REQ inter-arrival time Destination node selection
10
=2
1
1
Value Poisson at every router 200 time units, exponentially distributed Fixed 50,000 units 50,000 to 300,000 5% of the link capacity Computed using RV and STT Randomly distributed
P Pj j D ABD = P jP j na
i=1 na
i=1
Pi
j =1
ij
(4)
i
where na is total number of accepted requests, Pi is the primary path routed for request i, jPi j represents the length of the path and Dij is the actual backtracking distance to backup the j -th link of the primary path Pi . ABD measures the average backtracking distance among all the backup paths. Clearly, it captures expected case of backtracking and therefore, restoration latency.
Table 7.1 shows the parameters used in the experiments. Call holding time is how long each request lasts. Note that in reality, request load at various nodes may not be random and the amount of requests between certain node pairs may be dis-proportional. However, no real life call traffic data sets are currently available in public domain, therefore we use Poisson distribution to model request arrival 9
Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
7.2.2 Bandwidth Sharing Performance
1.4
EWSPF SPF D=0 D=1 D=∞
1.2
We used the following performance metrics in our evaluation [13]: (1) Fraction Rejected (FR): is the fraction of requests that are dropped during each simulation run. (2) Total bandwidth consumed: is the total bandwidth consumed by each scheme including primary and backup reservations.
Average BKD
1 0.8 0.6 0.4 0.2
7.3 Simulation Results
0 50
In the following, we discuss variation of these performance metrics for different cases of D and request volumes.
100
150 200 Request Volume in 1000’s
250
300
Figure 6. Average backtracking distance (ABD)
7.3.1 Restoration Latency Performance 1 EWSPF SPF D=∞ D=1 D=0
0.9
result in higher restoration latency and also, poorer control on providing tighter bounds on it. On the contrary, with our ; , the ABD stays within the pre-set algorithms for D limits. Note that the ABD for EWSPF is about longer than that for D 1 for all request size. The reason is that 1 tries to reduce the backtracking our algorithm for D distance even if it does not place a bound.
0.8
=01
Fraction of Backup Paths
0.7
0.6
=
0.5
0.4
0.3
0.2
12%
=
0.1
0
1
2
3
4
5 6 7 Backtracking Distance
8
9
10
11
12
7.3.2 Bandwidth Sharing Performance
13
Figure 5. Histogram of backtracking distance (HBD)
0.35
EWSPF SPF D=0 D=1 D=∞
0.3 0.25
Total Bandwidth Consumed (in 1000’s)
0 −1
FR
0.2 0.15
Figure 5 shows the histogram of backtracking distance for all the schemes using the request volume 300,000 as a representative case. For schemes EWSPF, SPF and D 1, backtracking distance is not bounded and can range from 0 to 13. Approximately and of backup paths require backtracking distance and respectively. For schemes D and D that bound the backtracking distance, the backtracking distance stays within the preset limits. Figure 6 shows the average backtracking distance ABD for all the schemes. We see that the two metrics have sim1 ilar behavior. Note that the ABD for EWSPF and D increases more rapidly than that for SPF. This happens because SPF has high rate of request rejection and utilizes network links rapidly. So it begins with shorter backup paths but as the links fill up, it can’t find longer feasible paths because a large subset of network links become infeasible due to lack of residual bandwidth. So with increased request rejection, the path length and therefore ABD metric remains flat. On the contrary, EWSPF exploits more bandwidth sharing and can route increasingly longer paths as load increases which results in higher ABD. Note that the case of D 1 results in a similar behavior. This clearly suggests that as the load on the network increases, higher ABD will
0.1
=
=0
=1
0.05 0
EWSPF SPF D=0 D=1 D=∞
1200 1100 1000 900 800 700 600 500 400 300 200
50
17% 8% 2 3
1300
100
150 200 250 Request Volume in 1000’s
(a) Fraction rejected (FR)
300
50
100
150 200 250 Request Volume in 1000’s
300
(b) Total bandwidth consumed
Figure 7. Bandwidth Sharing Performance
=
Fraction Rejected (FR): As expected, FR increases as the load or RV increases. All the sharing schemes are significantly better than SPF and can accept up to 22% more requests. This suggests backup bandwidth sharing enables much more efficient use of network resources than no-sharing case. There are two factors that impact FR. (1) Since end-to-end restoration schemes use one path to restore all the link failures in the primary path, fewer links need extra bandwidth reservation than D < 1. (2) Local restoration schemes can share backup bandwidth more aggressively than end-to-end schemes. This is because: in the end-to-end scheme, each link in the backup path has to assume that the maximum load in the primary link can
=
10 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
References
be backed up on that link (see Equation 2). On the other hand, in the local restoration schemes with post processing, each link only needs to backup the set of links whose backup path uses that link, that is, more bandwidth can be 1 are analogous in shared. Although EWSPF and D terms of unlimited backtracking, unlike the EWSPF case, backup paths for D 1 case can share primary path links. Therefore, FR for EWSPF should be higher than that of 1, (See Section 6). However, since the netscheme D works used in our simulation is relatively small, there are 1 can share primary not many cases where scheme D path links. Therefore, the FR for EWSPF and D 1 is indistinguishable. With respect to D and D , the first factor is more effective as many more links are needed for backup bandwidth reservation, this results in our algoand D rejects up to and more rithm for D requests respectively than EWSPF.
[1] G. Apostolopoulos, R. Guerin, and S.Kamat. QoS routing mechanisms and OSPF extensions LDP. work in progress, IETF draft, December 1998.
=
[2] D. et. al. Awduche. Requirements of Traffic Engineering over MPLS. RFC 2702, IETF, September 1999.
=
[3] B. Davie and Y. Rekhter. MPLS Technology and Applications. Morgan Kaufmann, San Francisco, CA, 2000.
=
=
=0
=0
=1
4%
[4] Loa Andersson (editor). PPVPN L2 Framework. Work in progress - Internet Draft, IETF, March 2002. draft-abobapppext-key-problem-01.txt.
= =1
[5] Jamoussi et al. Constraint-based LSP setup using LDP. RFC 3212, IETF, January 2002.
2%
[6] M. Charikar et. al. Approximation algorithms for directed Steiner problems. In Proceedings of the Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 192–200, January 1998.
Total Bandwidth Consumed: Figure 7-b shows the total bandwidth consumed for all the schemes. We see that even if SPF rejects up to 22% more request, it consumes much more bandwidth than the other schemes. For example, SPF when the reconsumes 28% more bandwidth than D quest volume is 150,000. Scheme D consumes 17% only conmore bandwidth than EWSPF. Scheme D sumes 3% more than EWSPF. This suggests, scheme D is a good tradeoff between restoration latency and network resource consumption.
=0 =0 =1
[7] P. Pan et al. Fast Reroute Extensions to RSVP-TE for LSP Tunnels. Internet Draft, draft-ietf-mpls-rsvp-lsp-fastreroute00.txt, pages 1–31, July 2002. [8] V. Sharma et al. Framework for MPLS-based Recovery. Work in progress - Internet Draft, IETF, January 2002. draftietf-mpls-recovery-frmwrk-03.txt.
=1
[9] M. Kodialam and T. V. Lakshman. Dynamic routing of bandwidth guaranteed tunnels with restoration. In Proceedings of IEEE INFOCOM, pages 902–911, 2000. [10] M. Kodialam and T. V. Lakshman. Minimum interference routing with applications to MPLS traffic engineering. In Proceedings of IEEE INFOCOM, pages 376–385, 2000.
8 Conclusions
[11] M. Kodialam and T.V. Lakshman. Dynamic routing of locally restorable bandwidth guaranteed tunnels using aggregated link usage information. In Proceedings of IEEE INFOCOM, pages 884–893, 2001.
In this paper, we proposed a framework for provisioning bandwidth guaranteed paths with bounded restoration latency. We introduced the novel concept of backtracking distance D and considered three different cases: (1) no backtracking (D ), (2) limited backtracking (D k ), and 1). We used a link cost (3) unlimited backtracking (D model that captures bandwidth sharing among links using various types of available link state information. We first showed that joint optimization of primary and backup paths is NP-hard in all cases. We then considered two-step algorithms where primary path and backup paths are computed in two separate steps. Using our link cost model, we devised heuristic algorithms for each case.
=0
=
[12] Li L., M. Buddhikot, C. Chekuri, and K. Guo. Dynamic routing of bandwidth guaranteed paths with local restoration. Bell Labs Technical Memorandum, 2002.
=
[13] S. Norden, M.M. Buddhikot M. Waldvogel, and S. Suri. Routing bandwidth guaranteed paths with restoration in label switched networks. In Proceedings of IEEE ICNP, pages 71–79, 2001. [14] S. Ramanathan. Multicast tree generation in networks with asymmetric links. IEEE/ACM Transactions on Networking, 4(4):558–568, 1996. [15] E. Rosen and Y. Rekhter. BGP/MPLS VPNs. RFC 2547, IETF, March 1999. [16] J. W. Suurballe. Disjoint paths in a network. Networks, 4:125–145, 1974.
Out simulation study to characterize the sharing and restoration latency performance of our schemes shows that D provides best tradeoff between bandwidth usage and restoration latency. Since the fault information only needs to be propagated at most one hop in this case, restoration latency performance may be acceptable for most service needs.
=1
[17] H. Takahashi and A. Matsuyama. An approximate solution for the steiner problem in graphs. Math. Japonica, 24(6):573–577, 1980. [18] D. Zhou and S. Subramaniam. Survivability in optical networks. IEEE Network, 14(6):16–23, December 2000.
11 Proceedings of the 10 th IEEE International Conference on Network Protocols (ICNP’02) 1092-1648/02 $17.00 © 2002 IEEE
